A Comparative Study of Machine Learning and Deep Ensemble Architectures for Cardiovascular Disease Risk Prediction on Imbalanced Health-Indicator Data

Cardiovascular disease (CVD) remains the single largest contributor to global mortality, yet
early screening is often out of reach in resource-constrained settings where clinical testing is
costly or unavailable. This paper investigates whether inexpensive, self-reported health
indicators can support reliable automated CVD risk screening. Fifteen classifiers spanning
classical machine learning, deep neural architectures, and two purpose-built deep ensembles—
EnsCVDD-Net and BlCVDD-Net—are evaluated on the publicly available Heart Disease Health
Indicators dataset. To counter the strong class skew inherent in population-level survey data,
the pipeline couples point-biserial correlation-driven feature selection with Adaptive Synthetic
(ADASYN) minority oversampling. Across accuracy, precision, recall, and F1-score, a soft-voting
ensemble of heterogeneous base learners delivers the strongest overall performance (91.7%
accuracy, 0.918 F1), outperforming every individual model as well as the deep architectures by a
wide margin, while remaining light enough for real-time inference. The selected model is
packaged into a browser-based screening application that returns instant risk estimates from
user-entered indicators. The results suggest that, for tabular health-indicator data of this kind,
carefully balanced classical ensembles remain a stronger and more deployable choice than
deeper networks, and that accessible pre-clinical screening tools can be built entirely from
open-source components.